What we measured after moving 100B tokens/week to open-weight models (www.ito.ai)

Ito has successfully migrated its production QA agents from the costly GPT-5.3-codex to the more economical open-weight model, MiniMax M3, leading to impressive savings and performance improvements. The migration, which processed 100 billion tokens per week, resulted in a 55% reduction in median LLM costs per pipeline run, with overall expenses decreasing by about 28% while workloads increased by 8%. The analysis reveals the significant variations cost-efficiency brings, particularly in context-heavy repositories where caching plays a crucial role in reducing expenses. This transition underscores the importance of not only the AI model but also the inference provider's infrastructure. The team highlighted that different providers could yield varied performance metrics due to differing caching behaviors and execution settings, suggesting that what is marketed as "open weights" may not behave uniformly across platforms. As a result, Ito recommends careful evaluation of providers, prioritization of compliance, and direct integrations wherever possible to truly optimize AI workloads. This case serves as a practical guide for organizations contemplating similar migrations, emphasizing that cost savings should be paired with thorough performance evaluations.